deeptrack.augmentations.FlipDiagonal assumes (X,Y,...) image format

[aarondomenzain]

The FlipDiagonal() method assumes that the first 2 axis correspond to spatial dimensions, i.e., (X,Y,C), as expressed in line 513:

This obviously creates a problem when training, since the usual convention is (N,C,X,Y). I am working on a fix for this, so it automatically detects which channels correspond to spatial dimensions. I am aware this fix must be also translated to update_properties() for supervised learning purposes.

[Pwhsky]

To make it channel-agnostic is a nice addition. Since we will soon have the option to use a torch backend, we will need to keep in mind (or leave it to the user) to manage the axis ordering as torch uses (B,C,X,Y).